Market · AI in practice
Meta launches Muse Ember for the repetitive work of agents
On November 9, 2026, Meta released Muse Ember, its smallest and most efficient model, at $0.10 per million input tokens and $0.50 per million output. Artificial Analysis puts it level with Claude Haiku 5.5 on its Intelligence Index at about a third of the cost per task.
What the sources report
On November 9, 2026, Meta introduced Muse Ember, a model designed for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. Pricing is $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens; above that, the rates rise to $0.30, $0.012 and $1.50. The model has a 1 million token context window, generates up to 512,000 output tokens and is available on the Meta Model API and in Muse Code. According to Meta, it is the first model trained using Muse Wonder Full.
In the independent evaluation by Artificial Analysis, Muse Ember scores 36 on the Intelligence Index, the same as Claude Haiku 5.5 at max effort, at a cost of $0.14 per task against $0.41. Meta suggests pairing Ember with Muse Wonder on the hardest problems, with Wonder planning and Ember handling the volume. Artificial Analysis names knowledge as the model's weak spot, with 21.1% accuracy on AA Omniscience, largely because it declines to answer 64% of the questions; as a result, its hallucination rate sits at 17.4%, half that of Claude Haiku 5.5.
My take
Reconciliation, document reading, data extraction and payment routing repeat all day in a financial operation, and that is the layer where cost per task decides whether an agent pays for itself. An executor at $0.14 per task on the Artificial Analysis index, working under a larger model that plans, changes the math for anyone currently using the same large model for everything. One more detail from the independent evaluation counts in its favor in finance: Ember would rather not answer than make something up, which holds hallucination at 17.4%. That caution has value in financial document processing, as long as the workflow design says where each declined case goes.
Sources
- https://research.meta.ai/blog/muse-ember (opens in a new tab) Meta
- https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/ (opens in a new tab) TechCrunch
- https://x.com/ArtificialAnlys/status/2119854348587123007 (opens in a new tab) Artificial Analysis (X)